Prof. Onur Mutlu Carnegie Mellon University 10/24/2012

Prof. Onur Mutlu Carnegie Mellon University 10/24/2012
Fall 2012 Parallel Computer Architecture Lecture 20: Speculation+Interconnects III Prof. Onur Mutlu Carnegie Mellon University 10/24/2012

New Review Assignments
Due: Sunday, October 28, 11:59pm. Das et al., “Aergia: Exploiting Packet Latency Slack in On-Chip Networks,” ISCA 2010. Dennis and Misunas, “A Preliminary Architecture for a Basic Data Flow Processor,” ISCA 1974. Due: Tuesday, October 30, 11:59pm. Arvind and Nikhil, “Executing a Program on the MIT Tagged-Token Dataflow Architecture,” IEEE TC 1990.

Due in the Future Dataflow Restricted Dataflow
Gurd et al., “The Manchester prototype dataflow computer,” CACM 1985. Lee and Hurson, “Dataflow Architectures and Multithreading,” IEEE Computer 1994. Restricted Dataflow Patt et al., “HPS, a new microarchitecture: rationale and introduction,” MICRO 1985. Patt et al., “Critical issues regarding HPS, a high performance microarchitecture,” MICRO 1985.

Project Milestone I Presentations (I)
When: October 26, in class Format: 9-min presentation per group, 2-min Q&A, 1-min grace period What to present: The problem you are solving + your goal Your solution ideas + strengths and weaknesses Your methodology to test your ideas Concrete mechanisms you have implemented so far Concrete results you have so far What will you do next? What hypotheses you have for future? How close were you to your target?

Project Milestone I Presentations (I)
You can update your slides Send them to me and Han by 2:30pm on Oct 26, Friday Make a lot of progress and find breakthroughs Example milestone presentations:

Last Few Lectures Speculation in Parallel Machines
Interconnection Networks Guest Lecture: Adam From, ARM

Today Transactional Memory (brief) Interconnect wrap-up

Review: Speculation to Improve Parallel Programs
Goal: reduce the impact of serializing bottlenecks Improve performance Improve programming ease Examples Herlihy and Moss, “Transactional Memory: Architectural Support for Lock-Free Data Structures,” ISCA 1993. Rajwar and Goodman, “Speculative Lock Elision: Enabling Highly Concurrent Multithreaded Execution,” MICRO 2001. Martinez and Torrellas, “Speculative Synchronization: Applying Thread-Level Speculation to Explicitly Parallel Applications,” ASPLOS 2002. Rajwar and Goodman, ”Transactional lock-free execution of lock-based programs,” ASPLOS 2002.

Review: Speculative Lock Elision
Many programs use locks for synchronization Many locks are not necessary Stores occur infrequently during execution Updates can occur to disjoint parts of the data structure Idea: Speculatively assume lock is not necessary and execute critical section without acquiring the lock Check for conflicts within the critical section Roll back if assumption is incorrect Rajwar and Goodman, “Speculative Lock Elision: Enabling Highly Concurrent Multithreaded Execution,” MICRO 2001.

Review: Dynamically Unnecessary Synchronization

Transactional Memory

Transactional Memory Idea: Programmer specifies code to be executed atomically as transactions. Hardware/software guarantees atomicity for transactions. Motivated by difficulty of lock-based programming Motivated by lack of concurrency (performance issues) in blocking synchronization (or “pessimistic concurrency”)

Locking Issues Locks: objects only one thread can hold at a time
Organization: lock for each shared structure Usage: (block)  acquire  access  release Correctness issues Under-locking  data races Acquires in different orders  deadlock Performance issues Conservative serialization Overhead of acquiring Difficult to find right granularity Blocking

Locks vs. Transactions Locks  pessimistic concurrency
Transactions  optimistic concurrency Lock issues: Under-locking  data races Deadlock due to lock ordering Blocking synchronization Conservative serialization How transactions help: Simpler interface/reasoning No ordering Nonblocking (Abort on conflict) Serialization only on conflicts

Transactional Memory Transactional Memory (TM) allows arbitrary multiple memory locations to be updated atomically (all or none) Basic Mechanisms: Isolation and conflict management: Track read/writes per transaction, detect when a conflict occurs between transactions Version management: Record new/old values (where?) Atomicity: Commit new values or abort back to old values  all or none semantics of a transaction Issues the same as other speculative parallelization schemes Logging/buffering Conflict detection Abort/rollback Commit

Four Issues in Transactional Memory
How to deal with unavailable values: predict vs. wait How to deal with speculative updates: logging/buffering How to detect conflicts: lazy vs. eager How and when to abort/rollback or commit

Many Variations of TM Software Hardware Hybrid HW/SW
High performance overhead, but no virtualization issues Hardware What if buffering is not enough? Context switches, I/O within transactions? Need support for virtualization Hybrid HW/SW Switch to SW to handle large transactions and buffer overflows

Initial TM Ideas Load Linked Store Conditional Operations
Lock-free atomic update of a single cache line Used to implement non-blocking synchronization Alpha, MIPS, ARM, PowerPC Load-linked returns current value of a location A subsequent store-conditional to the same memory location will store a new value only if no updates have occurred to the location Herlihy and Moss, ISCA 1993 Instructions explicitly identify transactional loads and stores Used dedicated transaction cache Size of transactions limited to transaction cache

Herlihy and Moss, ISCA 1993

Current Implementations of TM/SLE
Sun ROCK Dice et al., “Early Experience with a Commercial Hardware Transactional Memory Implementation,” ASPLOS 2009. IBM Blue Gene Wang et al., “Evaluation of Blue Gene/Q Hardware Support for Transactional Memories,” PACT 2012. IBM System z: Two types of transactions Best effort transactions: Programmer responsible for aborts Guaranteed transactions are subject to many limitations Intel Haswell

Some TM Research Issues
How to virtualize transactions (without much complexity) Ensure long transactions execute correctly In the presence of context switches, paging Handling I/O within transactions No problem with locks Semantics of nested transactions (more of a language/programming research topic) Does TM increase programmer productivity? Does the programmer need to optimize transactions?

Interconnects III: Review and Wrap-Up

Last Lectures Interconnection Networks Introduction & Terminology
Topology Buffering and Flow control Routing Router design Network performance metrics On-chip vs. off-chip differences

Some Questions What are the possible ways of handling contention in a router? What is head-of-line blocking? What is a non-minimal routing algorithm? What is the difference between deterministic, oblivious, and adaptive routing algorithms? What routing algorithms need to worry about deadlock? What routing algorithms need to worry about livelock? How to handle deadlock? How to handle livelock? What is zero-load latency? What is saturation throughput? What is an application-aware packet scheduling algorithm?

Routing Mechanism Arithmetic Source Based Table Lookup Based
Simple arithmetic to determine route in regular topologies Dimension order routing in meshes/tori Source Based Source specifies output port for each switch in route + Simple switches no control state: strip output port off header - Large header Table Lookup Based Index into table for output port + Small header - More complex switches Any advantages of Table Lookup Based? Easier to support fault tolerance. E.g. can bypass faulty links

Routing Algorithm Types How to adapt
Deterministic: always choose the same path Oblivious: do not consider network state (e.g., random) Adaptive: adapt to state of the network How to adapt Local/global feedback Minimal or non-minimal paths

Deterministic Routing
All packets between the same (source, dest) pair take the same path Dimension-order routing E.g., XY routing (used in Cray T3D, and many on-chip networks) First traverse dimension X, then traverse dimension Y + Simple + Deadlock freedom (no cycles in resource allocation) - Could lead to high contention - Does not exploit path diversity

Deadlock No forward progress
Caused by circular dependencies on resources Each packet waits for a buffer occupied by another packet downstream

Handling Deadlock Avoid cycles in routing
Dimension order routing Cannot build a circular dependency Restrict the “turns” each packet can take Avoid deadlock by adding virtual channels Detect and break deadlock Preemption of buffers

Turn Model to Avoid Deadlock
Idea Analyze directions in which packets can turn in the network Determine the cycles that such turns can form Prohibit just enough turns to break possible cycles Glass and Ni, “The Turn Model for Adaptive Routing,” ISCA 1992.

Valiant’s Algorithm An example of oblivious algorithm
Goal: Balance network load Idea: Randomly choose an intermediate destination, route to it first, then route from there to destination Between source-intermediate and intermediate-dest, can use dimension order routing + Randomizes/balances network load - Non minimal (packet latency can increase) Optimizations: Do this on high load Restrict the intermediate node to be close (in the same quadrant)

Adaptive Routing Minimal adaptive Non-minimal (fully) adaptive
Router uses network state (e.g., downstream buffer occupancy) to pick which “productive” output port to send a packet to Productive output port: port that gets the packet closer to its destination + Aware of local congestion - Minimality restricts achievable link utilization (load balance) Non-minimal (fully) adaptive “Misroute” packets to non-productive output ports based on network state + Can achieve better network utilization and load balance - Need to guarantee livelock freedom

More on Adaptive Routing
Can avoid faulty links/routers Idea: Route around faults + Deterministic routing cannot handle faulty components - Need to change the routing table to disable faulty routes - Assuming the faulty link/router is detected

Real On-Chip Network Designs
Tilera Tile64 and Tile100 Larrabee Cell

On-Chip vs. Off-Chip Differences
Advantages of on-chip Wires are “free” Can build highly connected networks with wide buses Low latency Can cross entire network in few clock cycles High Reliability Packets are not dropped and links rarely fail Disadvantages of on-chip Sharing resources with rest of components on chip Area Power Limited buffering available Not all topologies map well to 2D plane

Tilera Networks 2D Mesh Five networks Four packet switched
Dimension order routing, wormhole flow control TDN: Cache request packets MDN: Response packets IDN: I/O packets UDN: Core to core messaging One circuit switched STN: Low-latency, high-bandwidth static network Streaming data

We did not cover the following slides in lecture
We did not cover the following slides in lecture. These are for your preparation for the next lecture.

Research Topics in Interconnects
Plenty of topics in on-chip networks. Examples: Energy/power efficient/proportional design Reducing Complexity: Simplified router and protocol designs Adaptivity: Ability to adapt to different access patterns QoS and performance isolation Reducing and controlling interference, admission control Co-design of NoCs with other shared resources End-to-end performance, QoS, power/energy optimization Scalable topologies to many cores Fault tolerance Request prioritization, priority inversion, coherence, … New technologies (optical, 3D)

Packet Scheduling Which packet to choose for a given output port?
Router needs to prioritize between competing flits Which input port? Which virtual channel? Which application’s packet? Common strategies Round robin across virtual channels Oldest packet first (or an approximation) Prioritize some virtual channels over others Better policies in a multi-core environment Use application characteristics

Application-Aware Packet Scheduling
Das et al., “Application-Aware Prioritization Mechanisms for On-Chip Networks,” MICRO 2009.

The Problem: Packet Scheduling
App1 App2 App N App N-1 Network-on-Chip L2$ Bank mem cont Memory Controller P Accelerator Network-on-Chip Network-on-Chip is a critical resource shared by multiple applications

PE R PE R PE R PE Input Port with Buffers R PE R PE R PE R PE From East From West From North From South From PE VC 0 VC Identifier VC 1 VC 2 Control Logic Routing Unit ( RC ) VC Allocator VA Switch Allocator (SA) Crossbar ( 5 x ) To East To PE To West To North To South R PE R PE R PE R PE R PE R PE R PE R PE R Routers PE Processing Element (Cores, L2 Banks, Memory Controllers etc) Crossbar

VC Routing Unit ( RC ) VC Allocator VA Switch Allocator (SA) 1 2 From East From West From North From South From PE

Conceptual View VC 0 VC Routing Unit ( RC ) VC Allocator VA Switch 1 2 From East From West From North From South From PE From East VC 1 VC 2 Allocator (SA) From West From North From South From PE App1 App2 App3 App4 App5 App6 App7 App8

Conceptual View VC Routing Unit ( RC ) VC Allocator VA Switch Allocator (SA) 1 2 From East From West From North From South From PE VC 0 From East VC 1 VC 2 From West Scheduler From North From South Which packet to choose? From PE App1 App2 App3 App4 App5 App6 App7 App8

Existing scheduling policies Round Robin Age Problem 1: Local to a router Lead to contradictory decision making between routers: packets from one application may be prioritized at one router, to be delayed at next. Problem 2: Application oblivious Treat all applications packets equally But applications are heterogeneous Solution : Application-aware global scheduling policies.

Motivation: Stall Time Criticality
Applications are not homogenous Applications have different criticality with respect to the network Some applications are network latency sensitive Some applications are network latency tolerant Application’s Stall Time Criticality (STC) can be measured by its average network stall time per packet (i.e. NST/packet) Network Stall Time (NST) is number of cycles the processor stalls waiting for network transactions to complete

Motivation: Stall Time Criticality
Why applications have different network stall time criticality (STC)? Memory Level Parallelism (MLP) Lower MLP leads to higher STC Shortest Job First Principle (SJF) Lower network load leads to higher STC Average Memory Access Time Higher memory access time leads to higher STC

Application with high MLP
STC Principle 1 {MLP} Compute Observation 1: Packet Latency != Network Stall Time STALL of Red Packet = 0 STALL STALL Application with high MLP LATENCY LATENCY LATENCY

STC Principle 1 {MLP} Observation 1: Packet Latency != Network Stall Time Observation 2: A low MLP application’s packets have higher criticality than a high MLP application’s STALL of Red Packet = 0 STALL STALL Application with high MLP LATENCY LATENCY LATENCY Application with low MLP STALL STALL STALL LATENCY LATENCY LATENCY

STC Principle 2 {Shortest-Job-First}
Light Application Heavy Application Running ALONE Compute Baseline (RR) Scheduling 4X network slow down 1.3X network slow down SJF Scheduling 1.2X network slow down 1.6X network slow down Overall system throughput{weighted speedup} increases by 34%

Solution: Application-Aware Policies
Idea Identify stall time critical applications (i.e. network sensitive applications) and prioritize their packets in each router. Key components of scheduling policy: Application Ranking Packet Batching Propose low-hardware complexity solution

Component 1 : Ranking Ranking distinguishes applications based on Stall Time Criticality (STC) Periodically rank applications based on Stall Time Criticality (STC). Explored many heuristics for quantifying STC (Details & analysis in paper) Heuristic based on outermost private cache Misses Per Instruction (L1-MPI) is the most effective Low L1-MPI => high STC => higher rank Why Misses Per Instruction (L1-MPI)? Easy to Compute (low complexity) Stable Metric (unaffected by interference in network)

Component 1 : How to Rank? Execution time is divided into fixed “ranking intervals” Ranking interval is 350,000 cycles At the end of an interval, each core calculates their L1-MPI and sends it to the Central Decision Logic (CDL) CDL is located in the central node of mesh CDL forms a ranking order and sends back its rank to each core Two control packets per core every ranking interval Ranking order is a “partial order” Rank formation is not on the critical path Ranking interval is significantly longer than rank computation time Cores use older rank values until new ranking is available

Component 2: Batching Problem: Starvation Solution: Packet Batching
Prioritizing a higher ranked application can lead to starvation of lower ranked application Solution: Packet Batching Network packets are grouped into finite sized batches Packets of older batches are prioritized over younger batches Alternative batching policies explored in paper Time-Based Batching New batches are formed in a periodic, synchronous manner across all nodes in the network, every T cycles

Putting it all together
Before injecting a packet into the network, it is tagged by Batch ID (3 bits) Rank ID (3 bits) Three tier priority structure at routers Oldest batch first (prevent starvation) Highest rank first (maximize performance) Local Round-Robin (final tie breaker) Simple hardware support: priority arbiters Global coordinated scheduling Ranking order and batching order are same across all routers

STC Scheduling Example
8 7 Batch 2 Injection Cycles 6 5 Batching interval length = 3 cycles 4 Batch 1 Ranking order = 3 3 2 2 2 Batch 0 1 Core1 Core2 Core3 Packet Injection Order at Processor

Router 8 5 Batch 2 7 8 4 Injection Cycles 6 Scheduler 2 5 7 1 4 Batch 1 6 2 3 3 1 2 2 2 Batch 0 3 1 1 Applications

Router Round Robin Time 3 2 8 7 6 5 8 4 Scheduler 3 7 1 6 2 STALL CYCLES Avg RR 8 6 11 8.3 Age STC 2 3 2

Router Round Robin Time 5 4 3 1 2 2 3 2 8 7 6 5 Age 8 4 Time Scheduler 3 3 5 4 6 7 8 3 7 1 6 2 STALL CYCLES Avg RR 8 6 11 8.3 Age 4 7.0 STC 2 3 2

Ranking order STC Scheduling Example Router Round Robin Time 5 4 3 1 2 2 3 2 8 7 6 5 Age 8 4 Time Scheduler 1 2 2 2 3 3 5 4 6 7 8 3 STC 7 1 Time 3 5 4 6 7 8 6 2 STALL CYCLES Avg RR 8 6 11 8.3 Age 4 7.0 STC 1 3 5.0 2 3 2

Qualitative Comparison
Round Robin & Age Local and application oblivious Age is biased towards heavy applications heavy applications flood the network higher likelihood of an older packet being from heavy application Globally Synchronized Frames (GSF) [Lee et al., ISCA 2008] Provides bandwidth fairness at the expense of system performance Penalizes heavy and bursty applications Each application gets equal and fixed quota of flits (credits) in each batch. Heavy application quickly run out of credits after injecting into all active batches & stall till oldest batch completes and frees up fresh credits. Underutilization of network resources

System Performance STC provides 9.1% improvement in weighted speedup over the best existing policy{averaged across 96 workloads} Detailed case studies in the paper This graph compares the system performance provided by the four polices. Performance is normalized to that of the baseline RR scheduler. STC also improves system performance compared to the best previous scheduler, on all systems. It provides an 9.1% performance improvement. For the unfairness graph lower value is better

Slack-Driven Packet Scheduling
Das et al., “Aergia: Exploiting Packet Latency Slack in On-Chip Networks,” ISCA 2010.

Packet Scheduling in NoC
Existing scheduling policies Round robin Age Problem Treat all packets equally Application-oblivious Packets have different criticality Packet is critical if latency of a packet affects application’s performance Different criticality due to memory level parallelism (MLP) All packets are not the same…!!! There are two existing scheduling policies RR {explain} Age {explain} Lead to contradictory decision making between routers: packets from one application may be prioritized at one router, to be delayed at next. Stress bullets of application-oblivious Transition: Lets look at the motivation behind our proposed application-aware policies

MLP Principle Stall ( ) = 0 Stall Packet Latency != Network Stall Time
Compute Stall Latency ( ) Latency ( ) Latency ( ) with MLP cores issue several memory requests in parallel in the hope of overlapping future load misses with current load misses. First I show an application with high MLP, it does some computation and sends a bunch of packets(in this case three, green, blue, red). The core stalls for the packets to return. Note that blue packet’s latency is overlapped and hence it has much fewer stall cycles. Stall cycles of red packet is 0 because its latency is completely overlapped. Thus our first observation packet latency != NST and different packets have different stalll / impact on execution time Packet Latency != Network Stall Time Different Packets have different criticality due to MLP Criticality( ) > Criticality( )

Outline Introduction Aérgia Evaluation Conclusion Packet Scheduling
Memory Level Parallelism Aérgia Concept of Slack Estimating Slack Evaluation Conclusion

What is Aérgia? Aérgia is the spirit of laziness in Greek mythology
Some packets can afford to slack!

Slack of Packets What is slack of a packet?
Slack of a packet is number of cycles it can be delayed in a router without reducing application’s performance Local network slack Source of slack: Memory-Level Parallelism (MLP) Latency of an application’s packet hidden from application due to overlap with latency of pending cache miss requests Prioritize packets with lower slack

Concept of Slack Instruction Window Execution Time Network-on-Chip Latency ( ) Latency ( ) Load Miss Causes Slack Stall Compute Causes Load Miss We have two cores, core A colored green and core B colored red returns earlier than necessary Slack ( ) = Latency ( ) – Latency ( ) = 26 – 6 = 20 hops Packet( ) can be delayed for available slack cycles without reducing performance!

Prioritizing using Slack
Core A Packet Latency Slack 13 hops 0 hops 3 hops 10 hops 0 hops 4 hops 6 hops Causes Load Miss Slack( ) > Slack ( ) Load Miss Causes Interference at 3 hops Core B Now the packets B-1 and A-0 interfere in the network for 3 hops, and we can choose one over the other A-0 has slack of 0 hops and B-1 has a slack of 6 hops So we can Prioritize A-0 over B-1 to reduce stall cycles of Core-A without increasing stall cycles of Core-B Prioritize

Slack in Applications Non-critical critical
50% of packets have 350+ slack cycles critical 10% of packets have <50 slack cycles

Slack in Applications 68% of packets have zero slack cycles

Diversity in Slack

Diversity in Slack Slack varies between packets of different applications Slack varies between packets of a single application

Estimating Slack Priority
Slack (P) = Max (Latencies of P’s Predecessors) – Latency of P Predecessors(P) are the packets of outstanding cache miss requests when P is issued Packet latencies not known when issued Predicting latency of any packet Q Higher latency if Q corresponds to an L2 miss Higher latency if Q has to travel farther number of hops Difficult to predict packet latency Predicting Max (Predecessors’ Latencies) Number of predecessors Higher predecessor latency if predecessor is L2 miss Predicting latency of a packet P Higher latency if P corresponds to an L2 miss Higher latency if P has to travel farther number of hops

Slack of P = Maximum Predecessor Latency – Latency of P Slack(P) = PredL2: Set if any predecessor packet is servicing L2 miss MyL2: Set if P is NOT servicing an L2 miss HopEstimate: Max (# of hops of Predecessors) – hops of P PredL2 (2 bits) MyL2 (1 bit) HopEstimate

How to predict L2 hit or miss at core? Global Branch Predictor based L2 Miss Predictor Use Pattern History Table and 2-bit saturating counters Threshold based L2 Miss Predictor If #L2 misses in “M” misses >= “T” threshold then next load is a L2 miss. Number of miss predecessors? List of outstanding L2 Misses Hops estimate? Hops => ∆X + ∆ Y distance Use predecessor list to calculate slack hop estimate

Starvation Avoidance Problem: Starvation
Prioritizing packets can lead to starvation of lower priority packets Solution: Time-Based Packet Batching New batches are formed at every T cycles Packets of older batches are prioritized over younger batches

Putting it all together
Tag header of the packet with priority bits before injection Priority(P)? P’s batch (highest priority) P’s Slack Local Round-Robin (final tie breaker) Priority (P) = Batch (3 bits) PredL2 (2 bits) MyL2 (1 bit) HopEstimate (2 bits) Say “older batches are prioritized over younger batches”

Evaluation Methodology
64-core system x86 processor model based on Intel Pentium M 2 GHz processor, 128-entry instruction window 32KB private L1 and 1MB per core shared L2 caches, 32 miss buffers 4GB DRAM, 320 cycle access latency, 4 on-chip DRAM controllers Detailed Network-on-Chip model 2-stage routers (with speculation and look ahead routing) Wormhole switching (8 flit data packets) Virtual channel flow control (6 VCs, 5 flit buffer depth) 8x8 Mesh (128 bit bi-directional channels) Benchmarks Multiprogrammed scientific, server, desktop workloads (35 applications) 96 workload combinations We do our evaluations on a detailed cycle accurate CMP simulator with state-of-art NoC model. Our evaluations are across 35 diverse applications. Our results are for 96 workload combinations.

Qualitative Comparison
Round Robin & Age Local and application oblivious Age is biased towards heavy applications Globally Synchronized Frames (GSF) [Lee et al., ISCA 2008] Provides bandwidth fairness at the expense of system performance Penalizes heavy and bursty applications Application-Aware Prioritization Policies (SJF) [Das et al., MICRO 2009] Shortest-Job-First Principle Packet scheduling policies which prioritize network sensitive applications which inject lower load Age heavy applications flood the network higher likelihood of an older packet being from heavy application GSF Each application gets equal and fixed quota of flits (credits) in each batch. Heavy application quickly run out of credits after injecting into all active batches & stall till oldest batch completes and frees up fresh credits. Underutilization of network resources

System Performance SJF provides 8.9% improvement in weighted speedup
Aérgia improves system throughput by 10.3% Aérgia+SJF improves system throughput by 16.1%

Network Unfairness SJF does not imbalance network fairness
Aergia improves network unfairness by 1.5X SJF+Aergia improves network unfairness by 1.3X

Conclusions & Future Directions
Packets have different criticality, yet existing packet scheduling policies treat all packets equally We propose a new approach to packet scheduling in NoCs We define Slack as a key measure that characterizes the relative importance of a packet. We propose Aérgia a novel architecture to accelerate low slack critical packets Result Improves system performance: 16.1% Improves network fairness: 30.8%

Prof. Onur Mutlu Carnegie Mellon University 10/24/2012

Similar presentations

Presentation on theme: "Prof. Onur Mutlu Carnegie Mellon University 10/24/2012"— Presentation transcript:

Similar presentations

About project

Feedback

Log in

Auth with social network:

Prof. Onur Mutlu Carnegie Mellon University 10/24/2012

Similar presentations

Presentation on theme: "Prof. Onur Mutlu Carnegie Mellon University 10/24/2012"— Presentation transcript:

Similar presentations

About project

Feedback